Use per-run temp dirs for mounts instead of fixed /mnt paths

[yarikoptic-gitmate]

Summary

Changes all three backends (eval-under-nfs, eval-under-loop, eval-under-beegfs) to default to per-run mountpoints under $TMPDIR instead of fixed /mnt/<backend> paths. This eliminates collisions between concurrent runs, removes the need for pre-existing /mnt directories, and makes leftover mounts self-describing.

Key Changes

• Default mountpoint derivation: Each backend now derives MNT from a MNT_BASE created via mktemp -u, generating sibling paths like ~/.tmp/eval-under-nfs-B3UXj.orig (backing) and ~/.tmp/eval-under-nfs-B3UXj.nfs (mount). The old fixed defaults (/mnt/nfs, /mnt/loop, /mnt/beegfs) are no longer used unless explicitly passed via --mount-point or EVAL_UNDER_MOUNT.

• Selective teardown: Added MNT_CREATED flag to track whether the current run created the mountpoint. Teardown now only removes the mountpoint if this run created it, preserving pre-existing directories passed via --mount-point.

• Updated documentation:

  • Help text for --mount-point now explains the new default behavior with concrete examples
  • Added rationale for why the old fixed paths were problematic (collisions, root requirement, opacity)
  • Updated README with examples of the new per-run directory layout
  • Updated GOTCHAS.md to use generic <mount> placeholder instead of hardcoded /mnt/beegfs
  • Updated contributor guidelines in README to document the new pattern for future backends

• Consistent implementation across backends: All three backends follow the same pattern for consistency and maintainability.

Implementation Details

• MNT_BASE is created once per run using mktemp -u (reserves name without creating directory)
• Actual mountpoint directory is created by the backend's mount_*() function, which also sets MNT_CREATED=1 if the directory didn't pre-exist
• Explicit --mount-point still works as before and takes precedence over the derived default
• Backing state (NFS export dir, loop image, BeeGFS data root) remains on its own path to avoid self-referential mounts

https://claude.ai/code/session_011udwBohkmsizgX1nGUDUBG

[yarikoptic-gitmate]

CI: every red cell so far is the matrix baseline, not this change

Checked each failing cell of run 35807756537 against master's last scheduled run, 35590469985, which ran on eb99fc6 — this PR's base:

Cell	This PR	master baseline
Loop ext4 / git-annex test	fail	fail
Loop vfat / git-annex test	fail	fail
Loop vfat / git testsuite	fail	fail
Loop vfat / stress-ng	fail	fail
Loop vfat / pjdfstest	fail	fail
NFS (localhost) / pjdfstest	fail	fail
BeeGFS 7.4.6 / pjdfstest	fail	fail
BeeGFS 8.1.0 / pjdfstest	fail	fail

Not just the same conclusions — the same findings:

• NFS / pjdfstest reproduces the breakdown GOTCHAS records for that cell exactly: chown/00.t 106 of 1280 failing plus the "TODO passed" batch, chmod/00.t 1, unlink/14.t 1, across 238 files / 8798 assertions. The suite ran to completion on the relocated mount, which is the thing worth knowing here.
• Loop ext4 / git-annex test fails identically to master: 8 out of 13 tests failed, every one of them not enough free space, need ~78-82 MB more ... annex.diskreserve on get/copy against the 100 MB backing image.

The cells this change could plausibly have broken — where the mount moved out of /mnt/<backend> — are green: NFS / git-annex test, NFS / stress-ng, all three Loop ext4 suites, BeeGFS stress-ng on both versions.

Still running: NFS / git testsuite, BeeGFS 7.4.6 + 8.1.0 / git testsuite (green on master, so they are the ones left to confirm) and BeeGFS git-annex test on both versions (red on master). I'll follow up only if one of the baseline-green cells comes back red.

No re-run spent on the above: these are deterministic, documented filesystem findings reproduced assertion-for-assertion on the base branch, so a re-run has nothing to disambiguate.

---

Generated by Claude Code